Research on PM2.5 Spatiotemporal Forecasting Model Based on LSTM Neural Network

Fang Zhao, Ziyi Liang, Qiyan Zhang, Dewen Seng, Xiyuan ChenComputational 2021 · DOI

TL;DR

This paper introduces a novel spatiotemporal forecasting model for PM2.5 air pollution using Long Short-Term Memory (LSTM) neural networks. By integrating historical PM2.5 data, data from nearby monitoring stations, meteorological factors, and other pollutant concentrations, the model achieves improved accuracy in predicting future PM2.5 levels, crucial for public health and environmental management.

What Problem Does This Paper Address?

Air pollution, particularly fine particulate matter (PM2.5), poses significant threats to human health, contributing to cardiovascular and neurodegenerative diseases and reducing lifespan. While current air quality monitoring provides real-time data, predicting future air quality in advance is essential for timely warnings, public health interventions, and effective environmental management. However, air quality is a complex phenomenon influenced by numerous interconnected factors, including meteorological conditions (temperature, humidity, wind), traffic emissions, industrial activities, and the concentrations of other pollutants. These spatiotemporal dependencies make accurate forecasting a challenging task. Existing models often struggle to capture both the spatial relationships between different monitoring sites and the temporal evolution of pollutant concentrations.

Methods

The core of this research is a Spatiotemporal Long Short-Term Memory (ST-LSTM) model. The methodology involves several key steps:

  • Spatial Factor Extraction: To account for spatial dependencies, the authors identified the five nearest monitoring stations with the highest correlation in PM2.5 concentration sequences with the target station. This involves calculating distances between sites using the Haversine formula and measuring linear correlation using the Pearson correlation coefficient.
  • Data Fusion: Meteorological data (temperature, dew point, pressure, wind direction, wind speed) and other pollutant data (SO2, NO2, O3, CO) are merged with the PM2.5 concentration data from the target site and its correlated neighbors. All data is preprocessed, including outlier and missing value imputation, normalization, and scaling to [0,1].
  • Spatiotemporal Feature Extraction and Prediction: The fused, normalized time-series data is fed into an LSTM neural network. The LSTM architecture, known for its ability to handle long-term dependencies in sequential data, is employed to extract complex spatiotemporal features. The model is designed with multiple hidden LSTM layers and a fully connected output layer to predict PM2.5 concentrations for future time steps (t+1 to t+N).

The authors also conducted hyperparameter tuning, investigating the impact of the number of hidden layers and time steps on model performance, ultimately selecting three hidden layers and a time step of 14.

Datasets

The study utilizes hourly PM2.5 concentration data from 35 air quality monitoring stations in Beijing, collected from January 1, 2016, to December 31, 2017. Auxiliary data includes:

  • Meteorological Data: Obtained from the National Climate Data Center (NCDC).
  • Other Pollutant Data: SO2, NO2, O3, and CO concentrations, sourced from the Beijing Environmental Protection Testing Center.

The data was split into training and testing sets, with 67% used for testing and the remaining 33% for training.

Key Insights and Ideas

  • Spatial Correlation is Key: The paper strongly emphasizes that PM2.5 concentrations at a given site are significantly influenced by nearby sites. Identifying and incorporating data from highly correlated neighboring stations is crucial for improving prediction accuracy.
  • Multifaceted Input Improves Predictions: Merging diverse data sources—historical PM2.5, neighbor data, meteorological factors, and other pollutants—provides a richer context for the LSTM model, enabling it to capture more complex spatiotemporal relationships.
  • LSTM's Strength in Temporal Dependencies: The choice of LSTM is well-justified, as it effectively addresses the temporal nature of air pollution, where past conditions heavily influence future states.
  • R² as a Robust Indicator: The use of R² alongside MAE and RMSE provides a more comprehensive evaluation of the model's ability to explain the variance in the data, indicating how well it captures the underlying patterns.

Key Results

The proposed ST-LSTM model demonstrated superior performance compared to various baseline models (including non-deep learning models like LR, RF, SVM, ARMA, and several deep learning variants of LSTM).

  • Next-Hour Prediction: The ST-LSTM model achieved the best performance on the next-hour prediction task, as indicated by the lowest MAE and RMSE, and the highest R² (0.930) among all tested models (Table 3).
  • Longer-Term Prediction: For predictions up to 6 hours ahead, the ST-LSTM model consistently showed competitive results, often achieving the lowest MAE, particularly for the first few hours. While LSTM_NE sometimes outperformed it for longer horizons (4th-5th hour), the ST-LSTM's overall robustness across various time steps and its consideration of both meteorological and other pollutant factors make it a strong contender.
  • High Accuracy: The scatter plot (Figure 5) shows a strong visual correlation between observed and predicted PM2.5 values, with the R² value indicating that the model can capture 93% of the interpreted variance. The fit curve (Figure 6) further confirms the model's ability to track the trend of real PM2.5 values.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Single-City Focus: The current model is designed for predicting air quality at a single site within a single city. Future work could explore combining data from multiple sites across an entire city to achieve a comprehensive city-wide forecast.
  • Geographic Specificity: The model was evaluated solely on Beijing data. Further validation with data from other urban areas is needed to confirm its generalizability.
  • Additional Factors: The authors suggest incorporating more impact factors, such as traffic flow, to further enhance prediction accuracy and capture more nuanced changes in air quality.

Summary updated

Abstract

Accurate monitoring of air quality can no longer meet people's needs. People hope to predict air quality in advance and make timely warnings and defenses to minimize the threat to life. This paper proposed a new air quality spatiotemporal prediction model to predict future air quality and is based on a large number of environmental data and a long short-term memory (LSTM) neural network. In order to capture the spatial and temporal characteristics of the pollutant concentration data, the data of the five sites with the highest correlation of time-series concentration of PM2.5 (particles with aerodynamic diameter ≤2.5 mm) at the experimental site were first extracted, and the weather data and other pollutant data at the same time were merged in the next step, extracting advanced spatiotemporal features through long- and short-term memory neural networks. The model presented in this paper was compared with other baseline models on the hourly PM2.5 concentration data set collected at 35 air quality monitoring sites in Beijing from January 1, 2016, to December 31, 2017. The experimental results show that the performance of the proposed model is better than other baseline models.
BibTeX
@inproceedings{zhao2021research,
  author = {Fang Zhao and Ziyi Liang and Qiyan Zhang and Dewen Seng and Xiyuan Chen},
  title = {{Research on PM2.5 Spatiotemporal Forecasting Model Based on LSTM Neural Network}},
  booktitle = {Computational},
  year = {2021},
  doi = {10.1155/2021/1616806},
}
Figure